agent-project-sdlc 0.1.22 → 0.1.24

package/README.md

@@ -83,15 +83,23 @@ Release docs are current-state facts, not a version ledger. New release work sho
 
 The generic rule is that any workflow promoting a draft task into a formal `TASK-*` in `plan.yaml` must remove the source draft from its draft queue in the same state update. The formal task is then recovered only from `plan.yaml`; completed history lives in implementation docs, git, PR and CI records. The built-in Harness draft queue is currently `plan.draft.yaml.tasks[]`, which means unadopted development drafts only. `/devloop` treats the development queue as exhausted only when both `plan.yaml.tasks[]` and `plan.draft.yaml.tasks[]` have no executable task.
 
-Before development starts, `ARCHITECTING` can return to `REQUIREMENT_GATHERING` for PRD edits. The manager uses `python3 tools/transition.py --to REQUIREMENT_GATHERING`, the PM workflow updates the PRD through one `TASK-*`, then `validate-pm` and `python3 tools/transition.py --to ARCHITECTING` return the project to design. Requirement or design changes after `SPRINTING` use RFC recalibration; `SPRINTING`, `REVIEWING`, `TESTING` and `RELEASING` can enter the controlled interrupt with `python3 tools/transition.py --to RFC_RECALIBRATION`, then return to `SPRINTING` after `validate-rfc`.
+Phase routing is expressed as a lightweight explicit directed graph in `<harnessRoot>/pjsdlc_managed/policies/phase_contracts.yaml`: `phases` stores stable phase contracts, while `transitions` stores legal edges and small runtime effects such as setting or clearing `suspended_phase`. This makes normal advance, pre-development return, TESTING bugfix return, RFC interrupt/resume and BLOCKED resume rules consumable by both the transition helper and validators. It is intentionally not a heavy graph engine: no history graph, traversal framework, node/edge classes or visualizer are introduced; the goal is to reduce missed rules and drift.
 
-`validate-design` treats semantic slicing as a hard gate. Generated `overview.md` files do not count as deliverables, development draft tasks in `plan.draft.yaml` must reference existing tech plan slices through `docs.tech_plan`, multiple development draft tasks need distinct primary tech plan slices, and explicit AI provider/copilot, external-system, or compliance/permission/audit themes require dedicated architecture slices. Draft tasks with runnable boundaries must also include `self_test_contract`, backed by a `Development Self-Test Contract` section in the tech plan; the contract must include `module_key_test_path` from local start or invocation to all self-test scenarios completion, covering every runnable entry promised by the current task/module and its internal key paths.
+Migration cost is low for projects that use managed assets: run `npx sdlc-harness upgrade` to sync the new `phase_contracts.yaml` and `tools/transition.py`, or run `npx sdlc-harness sync` if only managed files need refreshing. `lifecycle.yaml` and `plan.yaml` do not need manual migration; old `allowed_next_phases` values are regenerated from the graph on the next transition. Projects with custom phase policies should convert node-local `next` / `returns` to top-level `transitions`, and add the `TESTING -> ARCHITECTING` / `TESTING -> SPRINTING` bugfix return edges when they want the new routing. If the new `validate-harness` reports missing `transitions`, run `upgrade` or `sync` before validating again.
 
-SPRINTING Definition of Done includes module-level runnable delivery boundaries. API, CLI, server route, service, agent, runtime, adapter, worker, provider, config-contract and fixture/live boundaries promised by a technical plan or task must be implemented or marked `BLOCKED` during development. Runtime/app/provider/live tasks must declare `evidence_level.required`, `target_runtime_environment` and `self_test_contract` in `plan.yaml`; every gate in `self_test_contract.required_gates` must also appear in task `required_gates`, and `self_test_contract.module_key_test_path` must describe the path from local start or invocation to all self-test scenarios completion, covering every runnable entry promised by the current task/module and its internal key paths. `deployed_runtime` cannot be closed by `unit`, `local_runtime`, `external_provider_live`, provider smoke, fake adapters or localhost smoke alone, and `business_handoff_ready` requires a Testing Handoff Contract. The current task implementation doc must include `Development Evidence` and a completed `Development Self-Test Report` with `Report Status: PASS | BLOCKED | IN_PROGRESS | STALE`, contract source, scenario results, executed gates, Module Key Test Path, actual evidence, missing/blockers and Testing Handoff Readiness; only `Report Status: PASS` with every scenario `PASS` can close a development task. The report proves module entry, core path, exit and minimal evidence; it is not a debug log, operator log, runbook or exploration history. Fallback/diagnostic detail belongs in `.docs/09_runbooks/**` appendices or git history. Module Key Test Path records actual entries, internal key paths, boundaries, checkpoints and observable completion evidence. Provider smoke, fixture smoke, fake adapters and one-shot smoke prove only local links; they do not by themselves prove application readiness. REVIEWING treats missing entry/exit, initialization, config contract, target runtime, evidence level or development evidence as blocking, and TESTING only exercises entrypoints that Review has confirmed as `PASS`; it must not add product runtime, bootstrap, provider adapter, deploy code or package runtime scripts.
+Before development starts, `ARCHITECTING` can return to `REQUIREMENT_GATHERING` for PRD edits. The manager uses `python3 tools/transition.py --to REQUIREMENT_GATHERING`, the PM workflow updates the PRD through one `TASK-*`, then `validate-pm` and `python3 tools/transition.py --to ARCHITECTING` return the project to design. Requirement, acceptance or product-boundary changes after `SPRINTING` use RFC recalibration; `SPRINTING`, `REVIEWING`, `TESTING` and `RELEASING` can enter the controlled interrupt with `python3 tools/transition.py --to RFC_RECALIBRATION`, then return to `SPRINTING` after `validate-rfc`.
 
-High-risk runtime/live/remote-operator tasks are resume-first. When the current SPRINTING task requires `external_provider_live`, `deployed_runtime` or `business_handoff_ready`, or its target runtime is `cloud_vm`, `managed_service`, `browser` or `worker`, `plan.yaml` must include top-level `resume_capsule` with the current state, canonical path, next step, blocker, last passed gate, do-not-retry list and recovery refs. Open task `working_notes` stays short, with a 5-8 item target and an 8 item validator limit. Long-term implementation facts stay in the implementation doc; operator paths, credential references and remote entrypoints live in `.docs/09_runbooks/**`; the implementation doc only keeps a short `Current Operator Path` with canonical operator path, runbook link, credential reference name, command/UI channel and do-not-retry summary. Evidence bodies live in an evidence index or external system; failed exploration stays in an exploration appendix. The Development Self-Test Report for these tasks must include a Gate Breakdown that separates local gate, cloud/service gate, executor/operator readiness and live smoke or handoff evidence.
+When TESTING finds a bug, first record `Bugfix Route` in `.docs/07_test/TEST_REPORT.md`, then let the manager choose the lightweight return. `bugfix_replan` uses `python3 tools/transition.py --to ARCHITECTING` when the technical plan, interface contract, task breakdown, Development Self-Test Contract or Module Key Test Graph must change. `bugfix_implementation_gap` uses `python3 tools/transition.py --to SPRINTING` only when the technical plan is still correct and implementation deviated from it. Requirement, acceptance or product-boundary changes still use RFC recalibration.
 
-`make validate-dev` and `npx sdlc-harness validate-dev` are in-development SPRINTING gates. They allow the current `current_task_id` open task to remain in `plan.yaml` while checking that it is a valid `phase: "SPRINTING"` task with `docs`, `allowed_paths`, `required_gates`, `acceptance_criteria`, `implementation_doc`, scoped dirty files, an empty `plan.draft.yaml` queue, runtime evidence task contract, `self_test_contract`, linked runnable-entry implementation docs, structured development evidence and a completed Development Self-Test Report. The report must include legal `Report Status` and Module Key Test Path so later agents can reuse the debug path from local entry to all self-test scenarios completion; that path is scoped to entries and internal key paths promised by the current task/module, not the whole system. `validate-dev` only passes completion-oriented dev evidence when `Report Status: PASS` and every scenario is `PASS`; `BLOCKED`, `IN_PROGRESS` and `STALE` reports may exist as recovery facts but cannot close the current development task. Page tasks need a dev server or page URL plus browser/Playwright/screenshot/equivalent interaction evidence; API/CLI/worker/service/agent/runtime tasks need a startup or invocation command, endpoint/health/status, and observable response/output/side effect. `validate-dev` checks content consistency and completeness between the report and current `self_test_contract`; it does not prove commands really executed in the current run. Agents must execute the current task `required_gates` before filling the report, and writing `PASS` without running those gates is an Agent execution violation. `make validate-current` and `/advance` are phase-exit gates; before moving to REVIEWING, the implementation commit and completion ledger must be done and no open task may remain.
+`validate-design` treats semantic slicing as a hard gate. Generated `overview.md` files do not count as deliverables, development draft tasks in `plan.draft.yaml` must reference existing tech plan slices through `docs.tech_plan`, multiple development draft tasks need distinct primary tech plan slices, and explicit AI provider/copilot, external-system, or compliance/permission/audit themes require dedicated architecture slices. Draft tasks with runnable boundaries must also include `self_test_contract`, backed by a `Development Self-Test Contract` section in the tech plan; the contract must include `module_key_test_path` from local start or invocation to all self-test scenarios completion, covering every runnable entry promised by the current task/module and its internal key paths. Complex or high-risk paths may set `graph_required: true` and provide `module_key_test_graph` to express entries, checkpoints, scenarios, exits and evidence refs as a lightweight DAG.
+
+SPRINTING Definition of Done includes module-level runnable delivery boundaries. API, CLI, server route, service, agent, runtime, adapter, worker, provider, config-contract and fixture/live boundaries promised by a technical plan or task must be implemented or marked `BLOCKED` during development. Runtime/app/provider/live tasks must declare `evidence_level.required`, `target_runtime_environment` and `self_test_contract` in `plan.yaml`; every gate in `self_test_contract.required_gates` must also appear in task `required_gates`, and `self_test_contract.module_key_test_path` must describe the path from local start or invocation to all self-test scenarios completion, covering every runnable entry promised by the current task/module and its internal key paths. Complex task `module_key_test_graph` is the canonical handoff path detail: it is a DAG instead of a tree because scenarios can share setup and converge on the same observable exit; it is not a heavy test execution graph and must not store traces, debug logs, operator logs, runbook bodies or evidence bodies. `deployed_runtime` cannot be closed by `unit`, `local_runtime`, `external_provider_live`, provider smoke, fake adapters or localhost smoke alone, and `business_handoff_ready` requires a Testing Handoff Contract. The current task implementation doc must include `Development Evidence` and a completed `Development Self-Test Report` with `Report Status: PASS | BLOCKED | IN_PROGRESS | STALE`, contract source, Module Application Entry, scenario results, executed gates, Module Key Test Path, Module Key Test Graph when required, Observable Exit, Current Blocker, Testing Handoff Readiness and Evidence Index Refs; only `Report Status: PASS` with every scenario `PASS` can close a development task. The report proves module entry, core path, exit and minimal evidence pointers; it is not a debug log, operator log, runbook, evidence dump or exploration history. Fallback/diagnostic detail belongs in `.docs/09_runbooks/**` evidence indexes, appendices or git history. The report must not use an `Actual Evidence` body field, should stay under 80 lines for ordinary tasks and under 120 lines for high-risk runtime tasks. Module Key Test Path records actual entries, internal key paths, boundaries, checkpoints and observable completion evidence. Provider smoke, fixture smoke, fake adapters and one-shot smoke prove only local links; they do not by themselves prove application readiness. REVIEWING treats missing entry/exit, initialization, config contract, target runtime, evidence level or development evidence as blocking, and TESTING only exercises entrypoints that Review has confirmed as `PASS`; complex paths are consumed through Module Key Test Graph, and TESTING must not add product runtime, bootstrap, provider adapter, deploy code or package runtime scripts.
+
+High-risk runtime/live/remote-operator tasks are resume-first. When the current SPRINTING task requires `external_provider_live`, `deployed_runtime` or `business_handoff_ready`, or its target runtime is `cloud_vm`, `managed_service`, `browser` or `worker`, `plan.yaml` must include top-level `resume_capsule` with the current state, canonical path, next step, blocker, last passed gate, do-not-retry list and recovery refs. Any judgment that changes the next action must be promoted to `resume_capsule.do_not_retry`, the runbook top-level `Hard Constraints`, or the short `Current Operator Path`; it cannot live only in evidence, notes, an appendix, or a long implementation doc. The validator scans `working_notes`, implementation docs and runbooks for session / QR / canonical path / do-not-retry judgments and fails when they are not promoted. Open task `working_notes` stays short, with a 5-8 item target and an 8 item validator limit. Long-term implementation facts stay in the implementation doc; operator paths, credential references and remote entrypoints live in `.docs/09_runbooks/**`; the implementation doc only keeps a short `Current Operator Path` with canonical operator path, runbook link, credential reference name, command/UI channel, hard constraints and do-not-retry summary. Evidence bodies live in an evidence index or external system and must not move into implementation-doc mainline sections such as `Evidence Dump`, `Operator Log`, `Failed Attempts`, or `Screenshot Index`; failed exploration stays in an exploration appendix. The Development Self-Test Report for these tasks must include a Gate Breakdown that separates local gate, cloud/service gate, executor/operator readiness and live smoke or handoff evidence.
+
+`make validate-dev` and `npx sdlc-harness validate-dev` are in-development SPRINTING gates. They allow the current `current_task_id` open task to remain in `plan.yaml` while checking that it is a valid `phase: "SPRINTING"` task with `docs`, `allowed_paths`, `required_gates`, `acceptance_criteria`, `implementation_doc`, scoped dirty files, an empty `plan.draft.yaml` queue, runtime evidence task contract, `self_test_contract`, linked runnable-entry implementation docs, structured development evidence and a completed Development Self-Test Report. The report must include legal `Report Status` and Module Key Test Path so later agents can reuse the debug path from local entry to all self-test scenarios completion; that path is scoped to entries and internal key paths promised by the current task/module, not the whole system. When `graph_required: true` or `module_key_test_graph` exists, validators check that the graph is a single-entry DAG, node and edge refs are valid, every scenario is reachable from the entry and can reach an observable exit, and `evidence_ref` is only a short pointer. `validate-dev` only passes completion-oriented dev evidence when `Report Status: PASS` and every scenario is `PASS`; `BLOCKED`, `IN_PROGRESS` and `STALE` reports may exist as recovery facts but cannot close the current development task. Page tasks need a dev server or page URL plus browser/Playwright/screenshot/equivalent interaction evidence; API/CLI/worker/service/agent/runtime tasks need a startup or invocation command, endpoint/health/status, and observable response/output/side effect. `validate-dev` checks content consistency and completeness between the report and current `self_test_contract`; it does not prove commands really executed in the current run. Agents must execute the current task `required_gates` before filling the report, and writing `PASS` without running those gates is an Agent execution violation. `make validate-current` and `/advance` are phase-exit gates; before moving to REVIEWING, the implementation commit and completion ledger must be done and no open task may remain.
+
+Migration cost for the lightweight DAG test path graph is zero to low. Existing projects, tasks and implementation docs that only use `module_key_test_path` remain valid; missing graphs are not retroactively rejected. Managed consumers can run `npx sdlc-harness upgrade`, or `npx sdlc-harness sync` when only managed files need refreshing. New high-risk or multi-scenario tasks will be prompted to generate `graph_required: true` and a graph skeleton. Existing high-risk tasks can be manually improved by splitting their current Module Key Test Path into a DAG. No automatic text-to-graph migration is provided because branches, checkpoints and observable exits require human or Agent judgment, and automatic conversion can invent false structure.
 
 `validate-test` keeps its command name as the TESTING phase gate. `.docs/07_test/TEST_STRATEGY.md` describes scope, environment, priority and execution strategy; `.docs/07_test/TEST_CASES.md` describes cases bound to real runnable entry/exit; `.docs/07_test/TEST_REPORT.md` only records executed TESTING evidence, test matrix, regression evidence, runnable entry/exit coverage, coverage gaps and final decision. `validate-test` only accepts `TEST_REPORT.md`; it no longer treats `TEST_PLAN.md` as a report fallback.
 

package/assets/agents/AGENTS_CORE.md

@@ -144,7 +144,7 @@ Strong success criteria 可以让你 independent loop。Weak criteria，例如
 6. 不要在当前 open task 的 `required_gates` 通过前把任务标记为 `done`。
 7. 代码 gate 通过后，更新相关 implementation doc 和 `.docs/INDEX.md`。
 8. `reviewer` 角色只读，不直接修改源码。
-9. 进入 `SPRINTING` 后的需求变更必须进入 RFC 工作流；`ARCHITECTING` 阶段发现 PRD 需要修改时，可以先回到 `REQUIREMENT_GATHERING`。
+9. 进入 `SPRINTING` 后的需求变更必须进入 RFC 工作流；`ARCHITECTING` 阶段发现 PRD 需要修改时，可以先回到 `REQUIREMENT_GATHERING`；`TESTING` 阶段发现 bug 时，先在 `TEST_REPORT.md` 判定 `Bugfix Route`，再轻量回到 `ARCHITECTING` 或 `SPRINTING`。
 10. task/release 历史动作记录使用 git commit、tag 或外部 release 系统，不维护 `<harnessRoot>/archive/` 常规归档。
 11. 在 `SPRINTING` 阶段，task 完成闭环必须先创建 task implementation commit，再提交移除该 task 后的 task completion ledger commit；如果没有 remote/upstream、权限或凭证导致无法 push，不要开始下一个 task，先报告 blocker。
 12. 文档 slice 发生变化后，运行 `make docs-overview` 刷新对应 `overview.md`。
@@ -168,6 +168,7 @@ Strong success criteria 可以让你 independent loop。Weak criteria，例如
 - “开始开发 / 做当前任务 / 做下一个任务” → 等价 `/dev`。
 - “开始循环：写任务，执行任务 / 把开发循环跑完” → 等价 `/devloop`。
 - “跑测试 / 验证一下” → 运行当前 task 或阶段对应 gate。
+- “测试发现 bug / 回去修 / 修测试问题” → 如果当前是 `TESTING`，先读取 `TEST_REPORT.md#Bugfix Route`；`bugfix_replan` 回 `ARCHITECTING` 修改技术方案后再开发，`bugfix_implementation_gap` 回 `SPRINTING` 补实现任务，需求或验收变化走 RFC。
 - 每个阶段任务开始时，默认先做 parallel eligibility check；适合安全拆分时，主 Agent 创建或使用 `parallel_execution.trigger: "workflow_default"` 并调度 Codex native subagents。用户说“并行 / 多 agent / 多 worktree / parallel” → 使用 `trigger: "user_requested"` 强化该意图。
 - “准备 review / 帮我 review” → 进入只读 Review workflow。
 - “刷新文档总览 / 同步 overview” → 等价 `/overview`。
@@ -200,3 +201,8 @@ python3 tools/transition.py --to <PHASE>
 
 流转前先运行阶段 gate，通常使用 `make validate-current`，或使用
 `.codex/pjsdlc_managed/policies/phase_contracts.yaml` 中列出的具体 `make validate-*` 目标。
+
+`TESTING` 中发现 bug 时不要直接重试或改 runtime。先让测试报告记录 `Bugfix Route`：`bugfix_replan`
+表示回 `ARCHITECTING` 修 tech plan / task breakdown / handoff graph；`bugfix_implementation_gap`
+表示技术方案仍正确、只回 `SPRINTING` 补实现偏差；需求、验收标准或产品边界变化仍走
+`RFC_RECALIBRATION`。

package/assets/docs/README.md

@@ -102,15 +102,23 @@ Agent 会读取 `<harnessRoot>/state/lifecycle.yaml` 和 `<harnessRoot>/state/pl
 
 技术方案阶段需要产出 `plan.draft.yaml`，是为了解决跨阶段交接和当前执行队列纯净性的冲突。`ARCHITECTING` 必须在进入开发前证明方案可以拆成具体、可验证的开发单元，包括修改范围、gate、implementation doc 和执行顺序；但这些未来开发 task 如果直接进入 `plan.yaml`，会和当前架构阶段 task 混在一起，让阶段 gate 无法区分“架构任务未完成”和“下一阶段任务已预拆”。因此开发任务先作为 draft 暂存，进入 `SPRINTING` 后再逐个 promote 成正式 `TASK-*`。其它阶段默认根据上一阶段已经稳定的事实源即时创建当前阶段 task，只有当某个阶段也需要提前为后续阶段生成具体执行任务时，才应引入同类 draft queue。
 
-在尚未进入开发前，`ARCHITECTING` 可以回到 `REQUIREMENT_GATHERING` 修改 PRD：Manager 使用 `python3 tools/transition.py --to REQUIREMENT_GATHERING` 切回 PM/PRD 工作流，完成 PRD task 和 `validate-pm` 后，再用 `python3 tools/transition.py --to ARCHITECTING` 回到设计阶段。进入 `SPRINTING` 后的需求或设计变化走 RFC workflow；`SPRINTING`、`REVIEWING`、`TESTING` 和 `RELEASING` 都可以通过 `python3 tools/transition.py --to RFC_RECALIBRATION` 进入受控 RFC 中断，RFC 完成后回到 `SPRINTING` 重新完成开发自测和 handoff。
+阶段关系由 `<harnessRoot>/pjsdlc_managed/policies/phase_contracts.yaml` 中的轻量显式有向图表达：`phases` 保存稳定阶段 contract，`transitions` 保存合法流转边和少量效果，例如设置或清理 `suspended_phase`。这样做是为了让正常推进、开发前返回、TESTING bugfix return、RFC interrupt/resume 和 BLOCKED resume 都被 transition helper 与 validator 读取，避免规则埋在长文档或工具硬编码里。它不是重型图引擎，不保存历史、不做复杂遍历、不引入 node/edge class 或可视化；目标只是降低遗漏和漂移。
 
-`validate-design` 会把架构阶段的语义切片作为硬 gate：`overview.md` 不计入 deliverables，`plan.draft.yaml` 中每个开发 draft task 必须通过 `docs.tech_plan` 指向存在的 tech plan slice；多个开发 draft task 默认需要不同 primary tech plan slice。PRD、tech plan 或 draft task 明确出现 AI provider / copilot、外部系统边界、合规 / 权限 / 审计等横切主题时，也需要对应的专门 architecture slice。可运行边界类 draft task 还必须带 `self_test_contract`，并在 tech plan 中有 `Development Self-Test Contract`；合同必须记录 `module_key_test_path`，说明从本地启动或调用入口开始，到完成全部自测 scenario 的模块关键测试路径，并覆盖本 task / 本模块承诺的所有可运行入口和内部关键路径。
+迁移成本较低：对使用 managed assets 的项目，运行 `npx sdlc-harness upgrade` 即可同步新的 `phase_contracts.yaml` 和 `tools/transition.py`；也可以运行 `npx sdlc-harness sync` 只刷新 managed 文件。`lifecycle.yaml` 和 `plan.yaml` 不需要手动迁移，旧的 `allowed_next_phases` 会在下一次 `transition.py` 执行后按图重新生成。只有维护了自定义 phase policy 的项目需要把阶段内的 `next` / `returns` 转成 top-level `transitions`，并按需加入 `TESTING -> ARCHITECTING` / `TESTING -> SPRINTING` bugfix return edges；如果升级前直接运行新版 `validate-harness` 看到缺少 `transitions`，先执行 `upgrade` / `sync`。
 
-SPRINTING 的 Definition of Done 包含模块级可运行交付边界：技术方案或 task 承诺的 API、CLI、server route、service、agent、runtime、adapter、worker、provider、配置契约和 fixture/live 边界必须在开发阶段实现或明确 `BLOCKED`。runtime/app/provider/live 类 task 必须在 `plan.yaml` 声明 `evidence_level.required`、`target_runtime_environment` 和 `self_test_contract`；`self_test_contract.required_gates` 必须同步出现在 task `required_gates`，`self_test_contract.module_key_test_path` 必须描述从本地启动或调用入口开始，到完成全部自测 scenario 的模块关键测试路径，并覆盖本 task / 本模块承诺的所有可运行入口和内部关键路径。`deployed_runtime` 不能用 `unit`、`local_runtime`、`external_provider_live`、provider smoke、fake adapter 或 localhost smoke 单独关闭，`business_handoff_ready` 还必须有 Testing Handoff Contract。当前 task 的 implementation doc 还必须写入 `Development Evidence` 和 `Development Self-Test Report`，其中自测报告记录 `Report Status: PASS | BLOCKED | IN_PROGRESS | STALE`、contract source、scenario results、executed gates、Module Key Test Path、actual evidence、missing/blockers 和 Testing Handoff Readiness；只有 `Report Status: PASS` 且所有 scenario 为 `PASS` 才能关闭 development task。`Development Self-Test Report` 只证明模块入口、核心路径、出口和最小证据，不承担 debug log、operator log、runbook 或探索流水职责；fallback / diagnostic 最多一句总结，详细内容进入 `.docs/09_runbooks/**` appendix 或 git history。`Module Key Test Path` 必须记录实际入口、内部关键路径、关键边界、观察点和可观测完成证据。provider smoke、fixture smoke、fake adapter 或 one-shot smoke 只能证明局部链路，不能单独证明 application readiness。REVIEWING 会把缺少入口/出口、初始化、配置契约、目标运行环境、证据等级或开发自测证据作为阻断项；TESTING 只调用 Review 已确认 `PASS` 的既有入口做输入输出验证，不能新增 product runtime、bootstrap、provider adapter、deploy 或 package runtime script。
+在尚未进入开发前，`ARCHITECTING` 可以回到 `REQUIREMENT_GATHERING` 修改 PRD：Manager 使用 `python3 tools/transition.py --to REQUIREMENT_GATHERING` 切回 PM/PRD 工作流，完成 PRD task 和 `validate-pm` 后，再用 `python3 tools/transition.py --to ARCHITECTING` 回到设计阶段。进入 `SPRINTING` 后的需求、验收标准或产品边界变化走 RFC workflow；`SPRINTING`、`REVIEWING`、`TESTING` 和 `RELEASING` 都可以通过 `python3 tools/transition.py --to RFC_RECALIBRATION` 进入受控 RFC 中断，RFC 完成后回到 `SPRINTING` 重新完成开发自测和 handoff。
 
-复杂 runtime/live/remote-operator 任务采用 resume-first 分层：当当前 SPRINTING task 要求 `external_provider_live`、`deployed_runtime`、`business_handoff_ready`，或目标环境是 `cloud_vm`、`managed_service`、`browser`、`worker` 时，`plan.yaml` 顶层必须维护 `resume_capsule`，只保留当前状态、canonical path、下一步、blocker、last passed gate、do-not-retry 和 recovery refs；open task 的 `working_notes` 只保留恢复短备注，目标 5-8 条且 validator 上限 8 条。长期实现事实写 implementation doc；操作路径、凭证引用、远端入口写 `.docs/09_runbooks/**` runbook；implementation doc 只放短的 `Current Operator Path`，记录 canonical operator path、runbook link、credential reference name、command/UI channel 和 do-not-retry summary；证据正文只进入 evidence index 或外部证据系统；失败探索隔离到 exploration appendix。高风险 task 的 `Development Self-Test Report` 还必须有 `Gate Breakdown`，把 local gate、cloud/service gate、executor/operator readiness 和 live smoke / handoff 分开记录，不能只用一个 `validate-dev PASS` 覆盖全部进度。
+TESTING 阶段发现 bug 时，先在 `.docs/07_test/TEST_REPORT.md` 记录 `Bugfix Route`，再由 Manager 选择轻量 return：`bugfix_replan` 走 `python3 tools/transition.py --to ARCHITECTING`，用于技术方案、接口契约、任务拆分、Development Self-Test Contract 或 Module Key Test Graph 需要改；`bugfix_implementation_gap` 走 `python3 tools/transition.py --to SPRINTING`，只用于技术方案正确但实现偏离的修复。后者是保留口子，不是预期常态；需求、验收标准或产品边界变化仍走 RFC。
 
-`make validate-dev` / `npx sdlc-harness validate-dev` 是 SPRINTING 开发中 gate：当前 `current_task_id` 指向的 open task 可以继续留在 `plan.yaml`，validator 会检查它是否是合法 `phase: "SPRINTING"` task、是否具备 `docs`、`allowed_paths`、`required_gates`、`acceptance_criteria`、`implementation_doc`，并校验 dirty files、`plan.draft.yaml`、runtime evidence task contract、`self_test_contract`、implementation doc、结构化 `Development Evidence` 和 `Development Self-Test Report`。自测报告必须记录合法 `Report Status` 和 `Module Key Test Path`，便于后续 Agent 复用从本地入口到全部自测用例完成的 debug 路径；该路径只要求覆盖本 task / 本模块承诺范围内的可运行入口和内部关键路径，不要求覆盖全系统所有模块。`validate-dev` 只接受 `Report Status: PASS` 且所有 scenario 为 `PASS` 的完成态；`BLOCKED`、`IN_PROGRESS`、`STALE` 可以记录恢复事实，但不能关闭当前 development task。页面类证据需要 dev server/page URL 与 browser check；API/CLI/worker/service/agent/runtime 类证据需要 startup/invocation command、endpoint/health/status 与 response/output/side effect。`validate-dev` 只校验自测报告内容与当前 `self_test_contract` 的一致性和完整性，不证明命令在本轮真实执行；Agent 必须先实际运行 current task `required_gates` 后再填写 `Development Self-Test Report`，未执行 required gates 却写 `PASS` 属于 Agent execution violation。`make validate-current` / `/advance` 是阶段出口 gate；进入 REVIEWING 前仍必须先完成 implementation commit 和 completion ledger，把 open task 从 `plan.yaml` 移除。
+`validate-design` 会把架构阶段的语义切片作为硬 gate：`overview.md` 不计入 deliverables，`plan.draft.yaml` 中每个开发 draft task 必须通过 `docs.tech_plan` 指向存在的 tech plan slice；多个开发 draft task 默认需要不同 primary tech plan slice。PRD、tech plan 或 draft task 明确出现 AI provider / copilot、外部系统边界、合规 / 权限 / 审计等横切主题时，也需要对应的专门 architecture slice。可运行边界类 draft task 还必须带 `self_test_contract`，并在 tech plan 中有 `Development Self-Test Contract`；合同必须记录 `module_key_test_path`，说明从本地启动或调用入口开始，到完成全部自测 scenario 的模块关键测试路径，并覆盖本 task / 本模块承诺的所有可运行入口和内部关键路径。复杂或 high-risk 路径可设置 `graph_required: true` 并提供 `module_key_test_graph`，把入口、checkpoint、scenario、出口和 evidence refs 表达成轻量 DAG。
+
+SPRINTING 的 Definition of Done 包含模块级可运行交付边界：技术方案或 task 承诺的 API、CLI、server route、service、agent、runtime、adapter、worker、provider、配置契约和 fixture/live 边界必须在开发阶段实现或明确 `BLOCKED`。runtime/app/provider/live 类 task 必须在 `plan.yaml` 声明 `evidence_level.required`、`target_runtime_environment` 和 `self_test_contract`；`self_test_contract.required_gates` 必须同步出现在 task `required_gates`，`self_test_contract.module_key_test_path` 必须描述从本地启动或调用入口开始，到完成全部自测 scenario 的模块关键测试路径，并覆盖本 task / 本模块承诺的所有可运行入口和内部关键路径。复杂 task 的 `module_key_test_graph` 是 handoff path 的 canonical detail：它是 DAG 而不是树，因为多个 scenario 可能共享 setup、分支后汇合到同一 observable exit；它不是重型测试执行图，不保存 trace、debug log、operator log、runbook 正文或证据正文。`deployed_runtime` 不能用 `unit`、`local_runtime`、`external_provider_live`、provider smoke、fake adapter 或 localhost smoke 单独关闭，`business_handoff_ready` 还必须有 Testing Handoff Contract。当前 task 的 implementation doc 还必须写入 `Development Evidence` 和 `Development Self-Test Report`，其中自测报告记录 `Report Status: PASS | BLOCKED | IN_PROGRESS | STALE`、contract source、Module Application Entry、scenario results、executed gates、Module Key Test Path、必要时的 Module Key Test Graph、Observable Exit、Current Blocker、Testing Handoff Readiness 和 Evidence Index Refs；只有 `Report Status: PASS` 且所有 scenario 为 `PASS` 才能关闭 development task。`Development Self-Test Report` 只证明模块入口、核心路径、出口和最小证据指针，不承担 debug log、operator log、runbook、evidence dump 或探索流水职责；fallback / diagnostic 最多一句总结，详细内容进入 `.docs/09_runbooks/**` evidence index / appendix 或 git history；主报告不得使用 `Actual Evidence` 正文字段，普通报告目标不超过 80 行，high-risk 报告目标不超过 120 行。`Module Key Test Path` 必须记录实际入口、内部关键路径、关键边界、观察点和可观测完成证据。provider smoke、fixture smoke、fake adapter 或 one-shot smoke 只能证明局部链路，不能单独证明 application readiness。REVIEWING 会把缺少入口/出口、初始化、配置契约、目标运行环境、证据等级或开发自测证据作为阻断项；TESTING 只调用 Review 已确认 `PASS` 的既有入口做输入输出验证，复杂路径按 Module Key Test Graph 消费，不能新增 product runtime、bootstrap、provider adapter、deploy 或 package runtime script。
+
+复杂 runtime/live/remote-operator 任务采用 resume-first 分层：当当前 SPRINTING task 要求 `external_provider_live`、`deployed_runtime`、`business_handoff_ready`，或目标环境是 `cloud_vm`、`managed_service`、`browser`、`worker` 时，`plan.yaml` 顶层必须维护 `resume_capsule`，只保留当前状态、canonical path、下一步、blocker、last passed gate、do-not-retry 和 recovery refs；凡会改变下一步动作的判断，必须 promoted 到 `resume_capsule.do_not_retry`、runbook 顶部 `Hard Constraints` 或短 `Current Operator Path`，不能只埋在 evidence、notes、appendix 或长 implementation doc 中。validator 会扫描 `working_notes`、implementation doc 和 runbook 中的 session / QR / canonical path / do-not-retry 类关键判断，未 promoted 时 fail。open task 的 `working_notes` 只保留恢复短备注，目标 5-8 条且 validator 上限 8 条。长期实现事实写 implementation doc；操作路径、凭证引用、远端入口写 `.docs/09_runbooks/**` runbook；implementation doc 只放短的 `Current Operator Path`，记录 canonical operator path、runbook link、credential reference name、command/UI channel、hard constraints 和 do-not-retry summary；证据正文只进入 evidence index 或外部证据系统，不得转移成 implementation doc 主线的 `Evidence Dump`、`Operator Log`、`Failed Attempts` 或 `Screenshot Index` 等章节；失败探索隔离到 exploration appendix。高风险 task 的 `Development Self-Test Report` 还必须有 `Gate Breakdown`，把 local gate、cloud/service gate、executor/operator readiness 和 live smoke / handoff 分开记录，不能只用一个 `validate-dev PASS` 覆盖全部进度。
+
+`make validate-dev` / `npx sdlc-harness validate-dev` 是 SPRINTING 开发中 gate：当前 `current_task_id` 指向的 open task 可以继续留在 `plan.yaml`，validator 会检查它是否是合法 `phase: "SPRINTING"` task、是否具备 `docs`、`allowed_paths`、`required_gates`、`acceptance_criteria`、`implementation_doc`，并校验 dirty files、`plan.draft.yaml`、runtime evidence task contract、`self_test_contract`、implementation doc、结构化 `Development Evidence` 和 `Development Self-Test Report`。自测报告必须记录合法 `Report Status` 和 `Module Key Test Path`，便于后续 Agent 复用从本地入口到全部自测用例完成的 debug 路径；该路径只要求覆盖本 task / 本模块承诺范围内的可运行入口和内部关键路径，不要求覆盖全系统所有模块。若 `graph_required: true` 或存在 `module_key_test_graph`，validator 会校验它是单入口 DAG、节点和边引用有效、每个 scenario 可从 entry 到达并能走到 observable exit，且 `evidence_ref` 只是短指针。`validate-dev` 只接受 `Report Status: PASS` 且所有 scenario 为 `PASS` 的完成态；`BLOCKED`、`IN_PROGRESS`、`STALE` 可以记录恢复事实，但不能关闭当前 development task。页面类证据需要 dev server/page URL 与 browser check；API/CLI/worker/service/agent/runtime 类证据需要 startup/invocation command、endpoint/health/status 与 response/output/side effect。`validate-dev` 只校验自测报告内容与当前 `self_test_contract` 的一致性和完整性，不证明命令在本轮真实执行；Agent 必须先实际运行 current task `required_gates` 后再填写 `Development Self-Test Report`，未执行 required gates 却写 `PASS` 属于 Agent execution violation。`make validate-current` / `/advance` 是阶段出口 gate；进入 REVIEWING 前仍必须先完成 implementation commit 和 completion ledger，把 open task 从 `plan.yaml` 移除。
+
+轻量 DAG 测试路径图的迁移成本为零到低：旧项目、旧 task 和旧 implementation doc 只写 `module_key_test_path` 仍然有效；缺少 graph 不会被 retroactive fail。升级 managed consumers 时运行 `npx sdlc-harness upgrade`，只刷新 managed files 时运行 `npx sdlc-harness sync`。新 high-risk / multi-scenario task 会由更新后的 prompts 倾向生成 `graph_required: true` 和 `module_key_test_graph` skeleton。旧 high-risk task 如需提升交接质量，可手动把现有 `Module Key Test Path` 拆成 DAG；不提供自动 text-to-graph migration，因为分支、checkpoint 和 observable exit 需要人工或 Agent 判断，自动转换容易制造虚假结构。
 
 `validate-test` 仍然是 TESTING 阶段 gate 名称。`.docs/07_test/TEST_STRATEGY.md` 描述测试范围、环境、优先级和执行策略；`.docs/07_test/TEST_CASES.md` 描述绑定真实 runnable entry/exit 的测试用例；`.docs/07_test/TEST_REPORT.md` 只记录 TESTING 阶段实际执行后的 test matrix、regression evidence、runnable entry/exit coverage、coverage gaps 和 final decision。`validate-test` 只接受 `TEST_REPORT.md`，不会把 `TEST_PLAN.md` 当作 report fallback。
 
@@ -128,7 +136,7 @@ SPRINTING 的 Definition of Done 包含模块级可运行交付边界：技术
 
 `<harnessRoot>/skills/<name>/SKILL.md` 是 Harness 的 workflow skill 事实源，也是稳定的 hard file index。它有两种使用方式：
 
-- Harness soft index：`AGENTS.md` 要求 Agent 先读 lifecycle/plan，再按 `active_skill` 和 `phase_contracts.yaml` 读取对应 skill。
+- Harness soft index：`AGENTS.md` 要求 Agent 先读 lifecycle/plan，再按 `active_skill` 和 `phase_contracts.yaml` 的 phase graph 读取对应 skill 与合法下一步。
 - Native skill adapter：如果某个 Agent 支持 native skill registry，可以把这些 workflow skills 同步或安装到该 Agent 的原生 skill root，让“产品方案”“技术方案”“开发循环”等自然语言在首轮就有机会命中。
 
 只在 `AGENTS.md` 里声明 `<harnessRoot>/skills` 不等于 native skill 注册；它保证的是 Harness soft index。Native skill 是否首轮水合，取决于具体 Agent 客户端是否扫描这个目录，或是否使用了额外 adapter。

package/assets/policies/phase_contracts.yaml

@@ -8,7 +8,6 @@ phases:
     outputs:
       - "<harnessRoot>/state/lifecycle.yaml"
     gates: []
-    next: "REQUIREMENT_GATHERING"
 
   REQUIREMENT_GATHERING:
     goal: "收集需求并形成产品方案"
@@ -23,7 +22,6 @@ phases:
       - ".docs/INDEX.md"
     gates:
       - "make validate-pm"
-    next: "ARCHITECTING"
 
   ARCHITECTING:
     goal: "根据产品方案生成架构设计、技术方案和任务草案"
@@ -40,9 +38,6 @@ phases:
       - "<harnessRoot>/state/plan.draft.yaml"
     gates:
       - "make validate-design"
-    next: "SPRINTING"
-    returns:
-      - "REQUIREMENT_GATHERING"
 
   SPRINTING:
     goal: "按任务状态执行开发、消费已采用草案、开发验证、Development Evidence 和实现文档沉淀"
@@ -62,7 +57,6 @@ phases:
       - "<harnessRoot>/state/plan.draft.yaml"
     gates:
       - "make validate-dev"
-    next: "REVIEWING"
 
   REVIEWING:
     goal: "只读审查实现质量、需求一致性和架构风险"
@@ -79,7 +73,6 @@ phases:
       - ".docs/06_review/REVIEW_REPORT.md"
     gates:
       - "make validate-review"
-    next: "TESTING"
 
   TESTING:
     goal: "基于已交付 entry/exit 形成测试策略、测试用例、执行报告、回归证据和覆盖缺口结论"
@@ -100,7 +93,6 @@ phases:
       - "tests/"
     gates:
       - "make validate-test"
-    next: "RELEASING"
 
   RELEASING:
     goal: "当前发布状态、发布检查和回滚方案"
@@ -115,7 +107,6 @@ phases:
       - ".docs/08_release/CURRENT_RELEASE.md"
     gates:
       - "make validate-release"
-    next: "COMPLETED"
 
   COMPLETED:
     goal: "当前里程碑已完成"
@@ -126,7 +117,6 @@ phases:
     outputs:
       - "<harnessRoot>/state/lifecycle.yaml"
     gates: []
-    next: "IDLE"
 
   RFC_RECALIBRATION:
     goal: "处理需求变更、影响分析、测试事实源清理、局部补丁和任务回退"
@@ -145,7 +135,6 @@ phases:
       - ".docs/INDEX.md"
     gates:
       - "make validate-rfc"
-    next: "SPRINTING"
 
   BLOCKED:
     goal: "等待人工补充信息、授权或基础设施处理"
@@ -156,4 +145,139 @@ phases:
     outputs:
       - "<harnessRoot>/state/lifecycle.yaml"
     gates: []
-    next: "REQUIREMENT_GATHERING"
+
+transitions:
+  - from: "IDLE"
+    to: "REQUIREMENT_GATHERING"
+    trigger: "advance"
+    kind: "normal"
+  - from: "REQUIREMENT_GATHERING"
+    to: "ARCHITECTING"
+    trigger: "advance"
+    kind: "normal"
+  - from: "ARCHITECTING"
+    to: "SPRINTING"
+    trigger: "advance"
+    kind: "normal"
+  - from: "SPRINTING"
+    to: "REVIEWING"
+    trigger: "advance"
+    kind: "normal"
+  - from: "REVIEWING"
+    to: "TESTING"
+    trigger: "advance"
+    kind: "normal"
+  - from: "TESTING"
+    to: "RELEASING"
+    trigger: "advance"
+    kind: "normal"
+  - from: "TESTING"
+    to: "ARCHITECTING"
+    trigger: "bugfix_replan"
+    kind: "return"
+  - from: "TESTING"
+    to: "SPRINTING"
+    trigger: "bugfix_implementation_gap"
+    kind: "return"
+  - from: "RELEASING"
+    to: "COMPLETED"
+    trigger: "advance"
+    kind: "normal"
+  - from: "COMPLETED"
+    to: "IDLE"
+    trigger: "advance"
+    kind: "normal"
+  - from: "ARCHITECTING"
+    to: "REQUIREMENT_GATHERING"
+    trigger: "return_to_prd"
+    kind: "return"
+  - from: "SPRINTING"
+    to: "RFC_RECALIBRATION"
+    trigger: "requirement_change"
+    kind: "interrupt"
+    effects:
+      set_suspended_phase: true
+  - from: "REVIEWING"
+    to: "RFC_RECALIBRATION"
+    trigger: "requirement_change"
+    kind: "interrupt"
+    effects:
+      set_suspended_phase: true
+  - from: "TESTING"
+    to: "RFC_RECALIBRATION"
+    trigger: "requirement_change"
+    kind: "interrupt"
+    effects:
+      set_suspended_phase: true
+  - from: "RELEASING"
+    to: "RFC_RECALIBRATION"
+    trigger: "requirement_change"
+    kind: "interrupt"
+    effects:
+      set_suspended_phase: true
+  - from: "RFC_RECALIBRATION"
+    to: "SPRINTING"
+    trigger: "rfc_complete"
+    kind: "resume"
+    effects:
+      clear_suspended_phase: true
+  - from: "IDLE"
+    to: "BLOCKED"
+    trigger: "blocked"
+    kind: "interrupt"
+    effects:
+      set_suspended_phase: true
+  - from: "REQUIREMENT_GATHERING"
+    to: "BLOCKED"
+    trigger: "blocked"
+    kind: "interrupt"
+    effects:
+      set_suspended_phase: true
+  - from: "ARCHITECTING"
+    to: "BLOCKED"
+    trigger: "blocked"
+    kind: "interrupt"
+    effects:
+      set_suspended_phase: true
+  - from: "SPRINTING"
+    to: "BLOCKED"
+    trigger: "blocked"
+    kind: "interrupt"
+    effects:
+      set_suspended_phase: true
+  - from: "REVIEWING"
+    to: "BLOCKED"
+    trigger: "blocked"
+    kind: "interrupt"
+    effects:
+      set_suspended_phase: true
+  - from: "TESTING"
+    to: "BLOCKED"
+    trigger: "blocked"
+    kind: "interrupt"
+    effects:
+      set_suspended_phase: true
+  - from: "RELEASING"
+    to: "BLOCKED"
+    trigger: "blocked"
+    kind: "interrupt"
+    effects:
+      set_suspended_phase: true
+  - from: "COMPLETED"
+    to: "BLOCKED"
+    trigger: "blocked"
+    kind: "interrupt"
+    effects:
+      set_suspended_phase: true
+  - from: "RFC_RECALIBRATION"
+    to: "BLOCKED"
+    trigger: "blocked"
+    kind: "interrupt"
+    effects:
+      set_suspended_phase: true
+  - from: "BLOCKED"
+    to: "<suspended_phase>"
+    trigger: "resume"
+    kind: "resume"
+    effects:
+      clear_suspended_phase: true

package/assets/skills/pjsdlc_architect_design/SKILL.md

@@ -23,6 +23,8 @@ ADR 用来解决“后来的人只看到结果，看不到当年取舍”的问
 
 如果在 `ARCHITECTING` 中发现 PRD 缺失、验收标准不清或产品边界需要调整，且项目尚未进入 `SPRINTING`，不要用架构文档替代产品事实。先收尾或移除当前 open design task，再请 Manager 使用 `python3 tools/transition.py --to REQUIREMENT_GATHERING` 回到 PM/PRD 工作流修改 `.docs/01_product/**`。进入 `SPRINTING` 后的需求变化走 RFC workflow。
 
+如果是从 `TESTING` 通过 `bugfix_replan` 回到 `ARCHITECTING`，默认只修正测试报告证明有问题的 tech plan、interface contract、task breakdown、Development Self-Test Contract 或 Module Key Test Graph，并引用 `.docs/07_test/TEST_REPORT.md` 中的失败 scenario。不要把这条路径扩成全量重设计；如果 bugfix 需要修改 PRD、acceptance criteria 或产品边界，转入 `RFC_RECALIBRATION`。
+
 架构阶段默认先评估是否适合并行调研或方案拆解。适合时，主 Agent 使用 `parallel_execution.trigger: "workflow_default"` 和 `runtime.provider: "codex_native_subagents"` 调度 worker 分别做架构草稿、接口分析、风险清单或方案对比；用户明确要求并行时使用 `trigger: "user_requested"`。worker 不直接写最终 architecture、tech plan、ADR 或 `plan.draft.yaml`，最终事实源和任务草案由主 Agent 合成；不适合拆分时保持串行并记录原因。
 
 ## 输入
@@ -57,7 +59,8 @@ ADR 用来解决“后来的人只看到结果，看不到当年取舍”的问
 - 如果实现计划改变了已有模块边界，应更新相关 architecture slice，而不是只在 task 描述里补一句。
 - 只要技术方案或 draft task 出现 service、agent、runtime、worker、frontend app、provider/live integration 或外部可运行边界，task breakdown 必须包含最后一公里 runtime 初始化和 testing handoff 交付：目标运行环境、启动/部署或预览方式、health/readiness、smoke 输入输出、日志/错误证据、测试可调用入口和出口。
 - 这类开发 draft task 必须写入 `evidence_level.required`、`target_runtime_environment` 和 `self_test_contract`。`evidence_level.required` 只能使用 `unit`、`local_runtime`、`external_provider_live`、`deployed_runtime`、`business_handoff_ready`；`target_runtime_environment.kind` 只能使用 `local`、`ci`、`staging`、`cloud_vm`、`managed_service`、`browser`、`worker`、`not_applicable`。`self_test_contract` 的 `source` 必须引用当前 tech plan slice，`required_gates` 必须同步到 task `required_gates`，`scenarios[]` 至少覆盖一个可运行入口和可观测出口，`module_key_test_path` 必须描述从本地启动或调用入口开始，到完成所有自测 scenario 的模块关键测试路径，并覆盖本 task / 本模块承诺的所有可运行入口和内部关键路径。
-- 如果 draft task 属于 high-risk runtime/live/remote-operator 工作（`external_provider_live`、`deployed_runtime`、`business_handoff_ready`，或目标环境为 `cloud_vm`、`managed_service`、`browser`、`worker`），还必须预留恢复分层：`docs.runbook` 指向 `.docs/09_runbooks/**` 下的 runbook / evidence index / exploration appendix，`allowed_paths` 覆盖这些文件，acceptance criteria 要求 promote 后维护 `plan.yaml#resume_capsule`。runbook 写 canonical operator path，evidence index 只写证据指针，exploration appendix 隔离失败尝试；不要把这些内容塞进 implementation doc 主线。
+- 复杂或高风险自测路径要同时生成轻量 DAG skeleton：当 scenario >= 3、多分支、多入口、runtime/live/provider/browser/worker，或 Review/Testing 需要明确消费 checkpoint / exit / evidence refs 时，在 `self_test_contract` 设置 `graph_required: true` 并填写 `module_key_test_graph`。图只保存 handoff path 的 `entry`、`checkpoint`、`branch`、`scenario`、`observable_exit` 节点和边，以及短 `evidence_ref` 指针；不要把命令输出、截图过程、operator log、debug log、runbook 正文、失败探索或历史流水塞进图。普通单入口 / 少量 scenario task 保留短 `module_key_test_path` 即可。
+- 如果 draft task 属于 high-risk runtime/live/remote-operator 工作（`external_provider_live`、`deployed_runtime`、`business_handoff_ready`，或目标环境为 `cloud_vm`、`managed_service`、`browser`、`worker`），还必须预留恢复分层：`docs.runbook` 指向 `.docs/09_runbooks/**` 下的 runbook / evidence index / exploration appendix，`allowed_paths` 覆盖这些文件，acceptance criteria 要求 promote 后维护 `plan.yaml#resume_capsule`，并把会改变下一步动作的判断 promoted 到 `resume_capsule.do_not_retry` 或 runbook 顶部 `Hard Constraints`。runbook 写 canonical operator path 和 hard constraints，evidence index 只写证据指针，exploration appendix 隔离失败尝试；不要把这些内容塞进 implementation doc 主线。
 - 如果用户明确要求把既有完整技术方案文件切成多个 `.docs/03_tech_plan/` slices，先确认 replacement slices 覆盖原文件中仍有效的接口契约、数据模型、模块方案、任务组和 gate；切片完成并更新 `plan.draft.yaml` 引用、`.docs/INDEX.md`、刷新 `overview.md` 后，删除被替代的完整 tech plan file，避免同一事实由完整文件和 slices 双重保留。
 - 每次新增、拆分、合并或废弃 slice 后，都要更新 `.docs/INDEX.md`。
 
@@ -83,6 +86,7 @@ ADR 用来解决“后来的人只看到结果，看不到当年取舍”的问
 5. 风险或不清晰的问题按 `<harnessRoot>/pjsdlc_managed/policies/risk_matrix.yaml` 标记。
 6. 任务边界应足够小，能在一次设计执行内闭环；`result_docs` 应指向将被更新或新增的 architecture、tech plan、ADR 或 `plan.draft.yaml` 文件。
 7. `make validate-design` 是阶段出口 gate；如果还有 open `TASK-*` design task，不要请求进入 `SPRINTING`。
+8. 从 TESTING bugfix 回流的设计任务必须引用 `TEST_REPORT.md` 的 `Bugfix Route: bugfix_replan` 和失败 scenario，只补受影响的 tech plan / draft task，不修改产品事实。
 
 ## 完成检查
 
@@ -94,6 +98,8 @@ ADR 用来解决“后来的人只看到结果，看不到当年取舍”的问
 - [ ] `plan.draft.yaml` 中每个开发 draft task 已通过 `docs.tech_plan` 绑定到对应 tech plan slice。
 - [ ] 如果用户要求把完整技术方案切成 tech plan slices，已删除被替代的完整 tech plan file，并同步 `plan.draft.yaml` 引用。
 - [ ] task draft 字段完整且范围清晰；runtime/app/provider/live 类 task 已声明 `evidence_level`、`target_runtime_environment` 和 `self_test_contract`。
+- [ ] 复杂或 high-risk 自测路径已判断是否需要 `graph_required: true`；需要时已在 tech plan 和 draft task 中生成 `module_key_test_graph` skeleton。
+- [ ] 如果来自 TESTING bugfix 回流，已只修正 `bugfix_replan` 指向的技术方案或任务边界，并保留 `TEST_REPORT.md` 失败 scenario 引用。
 - [ ] `.docs/INDEX.md` 已链接新增产物。
 - [ ] 已运行 `make docs-overview` 刷新 `.docs/<stage>/overview.md`。
 - [ ] `make validate-design` 准备通过。

package/assets/skills/pjsdlc_dev_sprint/SKILL.md

@@ -17,16 +17,18 @@ description: Use during SPRINTING to execute one task from plan.yaml, respecting
 
 开发阶段的 Definition of Done 包含可运行的系统入口/出口。凡技术方案或 task 承诺 API、CLI、server route、service、agent、runtime、adapter、worker、provider、外部发送/写入执行器、配置契约或 live/fixture 双模式边界，当前实现必须提供对应入口、调用方式、初始化方式、输出/副作用边界和验证方式；如果真实入口/出口尚不可运行，不能把 task 当作完成，也不能把缺口留给 TESTING 补 runtime。runtime/app/provider/live 类 task 必须在 `plan.yaml` 声明 `evidence_level.required`、`target_runtime_environment` 和 `self_test_contract`，并按合同交付：`deployed_runtime` 不能用 `unit`、`local_runtime`、`external_provider_live`、provider smoke、fake adapter 或 localhost smoke 单独关闭；`business_handoff_ready` 必须提供 Testing Handoff Contract。Implementation doc 必须写明 `Runnable Entry/Exit`，并在 `Development Evidence` 中记录 `Evidence Level`、`Target Runtime Environment`、`Runnable Entry`、`Observable Exit`、`Client / Server Initialization`、`Config Contract`、`Testing Handoff Readiness`、`Known Missing Runtime Boundaries` 和 `Basic Self-test Evidence`；其中 `Basic Self-test Evidence` 应指向已执行的 `Development Self-Test Report`。确实不适用时也要显式写 `Not applicable` 和具体原因。provider smoke、fixture smoke、fake adapter 或 one-shot smoke 只能证明局部链路，不能单独证明 `Application readiness`。此时应保留或创建 `BLOCKED`/后续 dev task，或通过 RFC/ARCHITECTING 处理边界变更。
 
-`self_test_contract` 是开发阶段自测合同，由 ARCHITECTING 或 RFC_RECALIBRATION 先定义，SPRINTING 负责执行并在 implementation doc 填写 `Development Self-Test Report`。开发者不得在开发结束后用现有实现反推自测合同；如果合同缺失、入口不匹配、required gates 未同步或场景无法执行，要先回到 ARCHITECTING/RFC 或把 task 保持为 `BLOCKED`。自测报告不是 TESTING 阶段产物，也不是 debug log、operator log、runbook 或历史流水；它只证明当前模块级可运行交付边界已经能被 Review/Testing 消费。报告必须写 `Report Status: PASS | BLOCKED | IN_PROGRESS | STALE`，只有 `PASS` 且所有 scenario 都是 `PASS` 才能关闭当前 development task；`BLOCKED`、`IN_PROGRESS`、`STALE` 可以作为恢复事实存在，但不能作为交接通过。报告还必须记录 `Module Key Test Path`：从本地启动或调用入口开始，执行并完成 `self_test_contract` 中全部自测用例的模块关键测试路径。该路径应覆盖本 task / 本模块承诺的所有可运行入口，以及自测用例实际经过的内部关键路径、关键边界、观察点和可观测完成证据，供后续 Agent 复用和 debug。
+`self_test_contract` 是开发阶段自测合同，由 ARCHITECTING 或 RFC_RECALIBRATION 先定义，SPRINTING 负责执行并在 implementation doc 填写 `Development Self-Test Report`。开发者不得在开发结束后用现有实现反推自测合同；如果合同缺失、入口不匹配、required gates 未同步或场景无法执行，要先回到 ARCHITECTING/RFC 或把 task 保持为 `BLOCKED`。自测报告不是 TESTING 阶段产物，也不是 debug log、operator log、runbook 或历史流水；它只证明当前模块级可运行交付边界已经能被 Review/Testing 消费。报告必须写 `Report Status: PASS | BLOCKED | IN_PROGRESS | STALE`，只有 `PASS` 且所有 scenario 都是 `PASS` 才能关闭当前 development task；`BLOCKED`、`IN_PROGRESS`、`STALE` 可以作为恢复事实存在，但不能作为交接通过。报告还必须记录 `Module Key Test Path`：从本地启动或调用入口开始，执行并完成 `self_test_contract` 中全部自测用例的模块关键测试路径。该路径应覆盖本 task / 本模块承诺的所有可运行入口，以及自测用例实际经过的内部关键路径、关键边界、观察点和可观测完成证据，供后续 Agent 复用和 debug。如果合同包含 `graph_required: true` 或 `module_key_test_graph`，开发者还必须沿该 DAG 执行并在报告中写入实际 `Module Key Test Graph`：节点只记录 entry、checkpoint、branch、scenario、observable_exit 和 evidence pointer，边只记录路径流向；不要把图扩成 runtime 搭建、执行 trace、debug log 或证据正文。
 
-开发阶段交付包含两类产物：实现产物（代码、配置、脚本、测试等）和开发自测产物。`Development Self-Test Report` 是开发阶段产物，不是计划、模板或历史记录。若当前 task 或关联技术方案声明 `self_test_contract.status: "required"`，必须先逐条执行 `self_test_contract.scenarios[]` 和 `self_test_contract.required_gates`，再填写或更新 `Development Self-Test Report`。没有本轮执行过的 runnable entry、内部关键路径、observable exit / artifact / screenshot / response / log 等证据时，不得写 `PASS`，不得完成 task。fallback / diagnostic 在主报告最多一句总结，详细命令、截图过程、UI 操作细节和失败路径进入 `.docs/09_runbooks/**` exploration appendix 或 git history。
+开发阶段交付包含两类产物：实现产物（代码、配置、脚本、测试等）和开发自测产物。`Development Self-Test Report` 是开发阶段产物，不是计划、模板或历史记录。若当前 task 或关联技术方案声明 `self_test_contract.status: "required"`，必须先逐条执行 `self_test_contract.scenarios[]` 和 `self_test_contract.required_gates`，再填写或更新 `Development Self-Test Report`。没有本轮执行过的 runnable entry、内部关键路径、observable exit / artifact / screenshot / response / log 等证据时，不得写 `PASS`，不得完成 task。自测报告只保留交接卡字段：`Report Status`、`Contract Source`、`Module Application Entry`、`Module Key Test Path`、`Scenario Results`、`Executed Gates`、`Observable Exit`、`Current Blocker`、`Testing Handoff Readiness` 和 `Evidence Index Refs`；证据正文、长命令输出、截图过程和 UI 操作细节进入 evidence index、外部 artifact 或 `.docs/09_runbooks/**` exploration appendix。主报告目标几十行，fallback / diagnostic 最多一句总结，不写 `Actual Evidence` 正文字段。
 
-高风险 runtime/live/remote-operator task 必须维护恢复优先级。若 `evidence_level.required` 是 `external_provider_live`、`deployed_runtime`、`business_handoff_ready`，或 `target_runtime_environment.kind` 是 `cloud_vm`、`managed_service`、`browser`、`worker`，`plan.yaml` 顶层必须有 `resume_capsule`，并在路径选择结论变化时立即更新：`state`、`canonical_path`、`next_step`、`blocker`、`last_passed_gate`、`do_not_retry` 和 `recovery_refs`。`working_notes` 只保留短恢复备注，目标 5-8 条且不得超过 8 条；canonical operator path 写入 `.docs/09_runbooks/**` runbook，并在 implementation doc 写一个短的 `Current Operator Path` 链接 runbook、credential reference name、command/UI channel 和 do-not-retry summary。证据正文只在 evidence 文件或外部系统，失败探索写入 exploration appendix。不要把 A/B/C 路径探索流水混进 implementation doc 主线或 scenario evidence。
+高风险 runtime/live/remote-operator task 必须维护恢复优先级。若 `evidence_level.required` 是 `external_provider_live`、`deployed_runtime`、`business_handoff_ready`，或 `target_runtime_environment.kind` 是 `cloud_vm`、`managed_service`、`browser`、`worker`，`plan.yaml` 顶层必须有 `resume_capsule`，并在路径选择结论变化时立即更新：`state`、`canonical_path`、`next_step`、`blocker`、`last_passed_gate`、`do_not_retry` 和 `recovery_refs`。凡会改变下一步动作的判断，必须 promoted 到 `resume_capsule.do_not_retry`、runbook 顶部 `Hard Constraints` 或 implementation doc 的短 `Current Operator Path`，不能只埋在 evidence、notes、exploration appendix 或长 implementation doc 中。例如 PC 微信已确认登录后再次出现 QR 时，先判定 `rule_assumption_gap` vs `operator_induced_logout_or_session_reset`，不得直接进入重新扫码流程。validator 会扫描 `working_notes`、implementation doc 和 runbook 中的 session / QR / canonical path / do-not-retry 类关键判断；只在 notes/evidence 中出现而未 promoted 会 fail。`working_notes` 只保留短恢复备注，目标 5-8 条且不得超过 8 条；canonical operator path 写入 `.docs/09_runbooks/**` runbook，并在 implementation doc 写一个短的 `Current Operator Path` 链接 runbook、credential reference name、command/UI channel、hard constraints 和 do-not-retry summary。证据正文只在 evidence index 或外部系统，失败探索写入 exploration appendix。不要把 A/B/C 路径探索流水混进 implementation doc 主线或 scenario evidence。
 
 页面类任务在开发阶段必须启动 dev server 或等价预览入口，并用浏览器、Playwright、截图或等价方式验证页面可加载、主入口可访问、核心按钮/表单/跳转可用、没有明显报错或空白页。API/CLI/worker/RPA/service/agent/runtime 类任务必须记录实际启动或调用命令、endpoint、worker command、dry-run/live preflight、health/status 或 server action，以及可观察的 response、队列 item、审计日志、文件产物、发送结果、错误码或 PASS/BLOCKED 结果。
 
 `/dev` 和 `/devloop` 是开发阶段的两个入口。`/dev` 创建或选择下一个最小 `TASK-*` development task，设置 `phase: "SPRINTING"`，并只完成一个 task 闭环后停止。通用规则是从任何 draft queue promote 正式 `TASK-*` 时都必须同次消费源 draft；当前开发阶段的内置 draft queue 是 `plan.draft.yaml.tasks[]`，因此如果这个 task 来自 `plan.draft.yaml.tasks[]`，promote 时必须同次删除源 draft，避免已采用草案继续显示为 `pending`。`/devloop` 连续运行 `/dev`，直到 `plan.yaml.tasks[]` 和 `plan.draft.yaml.tasks[]` 都没有明确可创建/执行的任务，或遇到需求、架构、allowed_paths、gate、commit/push blocker。
 
+如果是从 `TESTING` 通过 `bugfix_implementation_gap` 直接回到 `SPRINTING`，先创建或选择一个最小 bugfix `TASK-*`，引用 `.docs/07_test/TEST_REPORT.md` 的 failing scenario、既有 tech plan slice 和相关 implementation doc。该路径只用于“技术方案正确但实现偏离”的修复；如果测试证明技术方案、接口契约、任务拆分或 handoff graph 需要修改，应回 `ARCHITECTING`；如果需求或验收标准变化，应走 RFC。
+
 实现时遵循小步闭环：先检查 `git status`，确认工作区没有未归属到当前 task 的脏变更；再定位相关代码和测试，做必要修改，运行 gate，修复失败，写入或更新相关 implementation doc 并刷新文档派生视图。直接运行 `make validate-dev` 或 `npx sdlc-harness validate-dev` 是开发中 gate，允许当前 `SPRINTING` task 仍然 open，并校验 `current_task_id`、task 合同、dirty files、draft queue 和 implementation doc。此时先不要从 `plan.yaml` 移除当前 task，要在当前 task 仍位于 `plan.yaml` 时创建 task implementation commit；随后再移除 task，创建 task completion ledger commit，并 push 两个 commit。`make validate-current` / `/advance` 是阶段出口 gate，必须在 open task 已移除后才通过。不要顺手重构、重排格式或处理无关问题；如果发现无关风险，只记录或报告。
 
 开发阶段默认先评估当前 task 是否能安全并行。适合时，主 Agent 创建 `parallel_execution.trigger: "workflow_default"` 合同，默认使用 `runtime_managed` + `runtime.provider: "codex_native_subagents"` 调度 worker；用户明确要求并行、多 agent 或多 worktree 时使用 `trigger: "user_requested"`。主 Agent 声明每个 worker 的 `owned_paths`、`forbidden_paths`、`expected_output` 和 `required_gates`；非 native fallback 写仓库时还要声明 `branch` 和 `worktree`。worker 可以在各自非空、互不重叠且属于当前 task `allowed_paths` 的 owned paths 内实现，但不得直接修改 `plan.yaml`、`lifecycle.yaml`、`.docs/INDEX.md`、overview 或最终 implementation doc。主 Agent 负责 review、merge/cherry-pick、运行总 gate、更新事实源和完成两段提交。不适合拆分时继续串行 `/dev` 或 `/devloop` 并记录原因。
@@ -62,7 +64,7 @@ description: Use during SPRINTING to execute one task from plan.yaml, respecting
 - task implementation commit 必须发生在 task 移除之前，避免实现变更和计划短期化混在同一个提交里。
 - task completion ledger commit 发生在 implementation commit 之后，只负责将该 task 从当前 `plan.yaml` 移除。
 - 一个开发 task 默认对应一个主要 implementation commit 和一个轻量 completion ledger commit。implementation commit message 应包含 task id，例如 `TASK-003: implement login rate limit`；push 成功前，不进入下一个 task。
-- 本 Skill 不直接重切 PRD 或 tech plan；如果发现上游语义边界错误，进入 `BLOCKED`、创建 RFC，或请求回到 `ARCHITECTING`。
+- 本 Skill 不直接重切 PRD 或 tech plan；如果发现上游语义边界错误，进入 `BLOCKED`、创建 RFC，或请求回到 `ARCHITECTING`。从 TESTING bugfix 回流的 `bugfix_implementation_gap` task 也不得重切技术方案；它必须引用 `TEST_REPORT.md` 的失败 scenario 和既有 tech plan，并只修实现偏差。
 - gate 通过后调用 `pjsdlc_implementation_doc`，由该 Skill 按真实实现更新或新增 `.docs/04_implementation/` 模块级 slice。
 - direct dev gate 与 phase-exit gate 语义不同：`validate-dev` 支持当前 open `SPRINTING` task；`validate-current` 在 `SPRINTING` 下仍会拒绝 open task，提示先完成 implementation commit 和 completion ledger。
 - 如果一个任务实际变成多个独立实现边界，应停止扩大范围，拆分后续任务或回到任务规划。
@@ -75,7 +77,7 @@ description: Use during SPRINTING to execute one task from plan.yaml, respecting
 每个 open task 都必须在 `plan.yaml` 中包含完整执行合同：
 
 1. `current_task_id` 指向正在执行的 open task。
-2. open task 直接声明 `phase: "SPRINTING"`、`docs`、`allowed_paths`、`required_gates`、`acceptance_criteria` 和 `implementation_doc`；runtime/app/provider/live 类 task 还必须声明 `evidence_level`、`target_runtime_environment` 和 `self_test_contract`。`self_test_contract.required_gates` 必须同步出现在 task `required_gates`，`self_test_contract.module_key_test_path` 必须描述从本地启动或调用入口开始，到完成所有自测 scenario 的模块关键测试路径，并覆盖本 task / 本模块承诺的所有可运行入口和内部关键路径。
+2. open task 直接声明 `phase: "SPRINTING"`、`docs`、`allowed_paths`、`required_gates`、`acceptance_criteria` 和 `implementation_doc`；runtime/app/provider/live 类 task 还必须声明 `evidence_level`、`target_runtime_environment` 和 `self_test_contract`。`self_test_contract.required_gates` 必须同步出现在 task `required_gates`，`self_test_contract.module_key_test_path` 必须描述从本地启动或调用入口开始，到完成所有自测 scenario 的模块关键测试路径，并覆盖本 task / 本模块承诺的所有可运行入口和内部关键路径；若 `graph_required: true`，`module_key_test_graph` 是复杂路径的 canonical detail，不能用长文字路径替代。
 3. 如果 open task 是由 `plan.draft.yaml.tasks[]` promote 而来，创建正式 `TASK-*` 和删除源 draft 必须发生在同一次状态更新中；正式 task 的恢复现场只保存在 `plan.yaml`。
 4. 任务执行中只保留恢复所需的简短 `working_notes`，目标 5-8 条且不得超过 8 条；high-risk runtime/live task 用 `resume_capsule` 保存恢复卡片，并链接 runbook / evidence index / exploration appendix。
 5. gate、implementation doc、`.docs/INDEX.md` 和 `overview.md` 完成后，在当前 task 仍位于 `plan.yaml` 时创建 task implementation commit。
@@ -113,10 +115,12 @@ done task 的执行流水不在当前 `plan.yaml` 长期保留，也不是默认
 - [ ] 当前任务仍然是单一清晰的执行单元。
 - [ ] 技术方案或 task 承诺的 API/CLI/adapter/worker/provider、配置契约、输出/副作用和 fixture/live 边界已可运行并写入 implementation doc，或已明确 `BLOCKED`/后续 dev task。
 - [ ] implementation doc `Development Evidence` 已记录 `Evidence Level`、`Target Runtime Environment`、`Runnable Entry`、`Observable Exit`、`Client / Server Initialization`、`Config Contract`、`Testing Handoff Readiness`、`Known Missing Runtime Boundaries`、`Basic Self-test Evidence`，或写明带原因的 `Not applicable`。
-- [ ] 如果当前 task 有 `self_test_contract.status: "required"`，已逐条执行当前 `self_test_contract.scenarios[]` 和 `self_test_contract.required_gates`，并在 implementation doc `Development Self-Test Report` 写入 `Report Status`、本轮 contract source、scenario results、executed gates、runnable entry、内部关键路径、observable exit/evidence、Module Key Test Path、missing/blockers 和 Testing Handoff Readiness，且 `Report Status: PASS`、所有 scenario 都是 `PASS`。
+- [ ] 如果当前 task 有 `self_test_contract.status: "required"`，已逐条执行当前 `self_test_contract.scenarios[]` 和 `self_test_contract.required_gates`，并在 implementation doc `Development Self-Test Report` 写入 `Report Status`、本轮 contract source、Module Application Entry、scenario results、executed gates、Module Key Test Path、Observable Exit、Current Blocker、Testing Handoff Readiness 和 Evidence Index Refs，且 `Report Status: PASS`、所有 scenario 都是 `PASS`。
+- [ ] 如果 `self_test_contract.graph_required: true` 或存在 `module_key_test_graph`，已在 `Development Self-Test Report` 写入实际 `Module Key Test Graph`，且 graph 覆盖入口、scenario、observable exit 和 evidence refs。
 - [ ] 如果当前 task 是 high-risk runtime/live/remote-operator 工作，`resume_capsule` 已更新为 5-8 条恢复事实，`recovery_refs` 链接 implementation doc 和 `.docs/09_runbooks/**` runbook/evidence，implementation doc 已填写短的 `Current Operator Path` 和分层 `Gate Breakdown`。
 - [ ] 如果 task 要求 `business_handoff_ready`，implementation doc 已写入 Testing Handoff Contract，包含入口、配置、初始化/health、输入样例、预期出口、清理方式和证据等级。
 - [ ] 如果当前 task 来自 `plan.draft.yaml.tasks[]`，源 draft 已在 promote 时从 draft 列表删除。
+- [ ] 如果当前 task 来自 TESTING bugfix 回流，已确认它是 `bugfix_implementation_gap`，并引用 `TEST_REPORT.md`、既有 tech plan 和 implementation doc。
 - [ ] implementation doc 已生成或更新，并反映相关模块的真实代码。
 - [ ] 如果启用了 `parallel_execution`，worker owned paths、forbidden paths、required gates 和主 Agent 集成结果已记录。
 - [ ] gate 结果已写入 implementation doc `Verification`，必要时当前 task `working_notes` 也记录了恢复现场所需的 gate evidence。

package/assets/skills/pjsdlc_implementation_doc/SKILL.md

@@ -21,7 +21,7 @@ implementation doc 只写长期实现事实，不写完整操作日记。对于
 
 如果模块包含或承诺可运行系统边界，implementation doc 必须记录 runnable entry/exit：API/CLI/server route/service/agent/runtime/adapter/worker/provider 的调用方式、初始化方式、配置契约、输入来源、输出或副作用、fixture/live 模式边界，以及哪些真实外部执行器尚未实现。还必须在 `Development Evidence` 中记录开发阶段实际验证过的 `Evidence Level`、`Target Runtime Environment`、`Runnable Entry`、`Observable Exit`、`Client / Server Initialization`、`Config Contract`、`Testing Handoff Readiness`、`Known Missing Runtime Boundaries` 和 `Basic Self-test Evidence`；`Basic Self-test Evidence` 应指向已执行的 `Development Self-Test Report`。确实没有应用入口时，`Not applicable` 必须写清原因。不能把未来才会实现的入口写成当前事实，不能把 provider smoke、fixture smoke、fake adapter 或 one-shot smoke 单独写成 application readiness。如果 task 要求 `business_handoff_ready`，还必须写 Testing Handoff Contract，包含入口、配置、初始化/health、输入样例、预期出口、清理/reset/幂等说明和证据等级。
 
-如果当前 task 有 `self_test_contract.status: "required"`，implementation doc 必须填写 `Development Self-Test Report`，把设计/RFC 阶段定义的自测合同执行完成：记录 `Report Status: PASS | BLOCKED | IN_PROGRESS | STALE`、contract source、每个 scenario 的结果、实际执行入口、实际出口、证据位置或命令输出、executed gates、Module Key Test Path、missing/blockers 和 Testing Handoff Readiness。`Development Self-Test Report` 不是 debug log、operator log、runbook 或历史流水；fallback / diagnostic 在主报告最多一句总结，详细命令、截图过程、UI 操作细节和失败路径进入 exploration appendix 或 git history。High-risk runtime/live task 还必须写 `Current Operator Path` 和 `Gate Breakdown`，把 canonical operator path、runbook link、credential reference name、command/UI channel、do-not-retry summary 以及 local gate、cloud/service gate、executor/operator readiness、live smoke / handoff 分层记录，不能只写一个大 `validate-dev PASS`。`Development Self-Test Report` 只能记录当前 task 本轮实际执行后的结果；不得用历史报告、模板字段、代码阅读或无关通用 gate 替代本轮 self-test scenario 执行。`Module Key Test Path` 必须说明从本地启动或调用入口开始，执行并完成 `self_test_contract` 中全部自测用例的模块关键测试路径。该路径应覆盖本 task / 本模块承诺的所有可运行入口，以及自测用例实际经过的内部关键路径、关键边界、观察点和可观测完成证据，供后续 Agent 复用和 debug。任何 scenario 非 `PASS`，或 `Report Status` 为 `BLOCKED`、`IN_PROGRESS`、`STALE` 时，不得把开发 task 写成完成。
+如果当前 task 有 `self_test_contract.status: "required"`，implementation doc 必须填写 `Development Self-Test Report`，把设计/RFC 阶段定义的自测合同执行完成：记录 `Report Status: PASS | BLOCKED | IN_PROGRESS | STALE`、contract source、`Module Application Entry`、每个 scenario 的结果、executed gates、`Module Key Test Path`、`Observable Exit`、`Current Blocker`、`Testing Handoff Readiness` 和 `Evidence Index Refs`。`Development Self-Test Report` 是几十行交接卡，不是 debug log、operator log、runbook、evidence dump 或历史流水；不要写 `Actual Evidence` 正文字段。fallback / diagnostic 在主报告最多一句总结，详细命令、截图过程、UI 操作细节和失败路径进入 evidence index、exploration appendix 或 git history。High-risk runtime/live task 还必须写 `Current Operator Path` 和 `Gate Breakdown`，把 canonical operator path、runbook link、credential reference name、command/UI channel、hard constraints、do-not-retry summary 以及 local gate、cloud/service gate、executor/operator readiness、live smoke / handoff 分层记录，不能只写一个大 `validate-dev PASS`。凡会改变下一步动作的判断，必须 promoted 到 `plan.yaml#resume_capsule.do_not_retry` 或 runbook 顶部 `Hard Constraints`，不能只留在 evidence 或 appendix。`Development Self-Test Report` 只能记录当前 task 本轮实际执行后的结果；不得用历史报告、模板字段、代码阅读或无关通用 gate 替代本轮 self-test scenario 执行。`Module Key Test Path` 必须说明从本地启动或调用入口开始，执行并完成 `self_test_contract` 中全部自测用例的模块关键测试路径。该路径应覆盖本 task / 本模块承诺的所有可运行入口，以及自测用例实际经过的内部关键路径、关键边界、观察点和可观测完成证据，供后续 Agent 复用和 debug。如果 task contract 设置 `graph_required: true` 或包含 `module_key_test_graph`，还必须记录实际 `Module Key Test Graph`，用轻量 DAG 展示 entry、checkpoint、branch、scenario、observable_exit 与 evidence refs；图只放路径骨架和证据指针，不放执行 trace、命令输出、截图过程、operator log、debug log、runbook 正文或失败探索。任何 scenario 非 `PASS`，或 `Report Status` 为 `BLOCKED`、`IN_PROGRESS`、`STALE` 时，不得把开发 task 写成完成。
 
 ## 输入
 
@@ -53,7 +53,7 @@ implementation doc 只写长期实现事实，不写完整操作日记。对于
 3. 与技术方案的偏移必须明确记录，即便该偏移是合理的。
 4. runnable entry/exit、配置契约和 fixture/live 边界必须记录当前事实；缺失项写入 `未覆盖（Not covered）` 或方案偏移。
 5. `Development Evidence` 必须包含 task 合同要求的证据等级、目标运行环境、实际可调用入口、可观察出口、初始化方式、配置契约、测试交接状态、缺失 runtime 边界和开发自测证据；页面类任务记录 dev server/page URL 与 browser check，API/CLI/worker/RPA/service/agent/runtime 类任务记录 startup/invocation command、endpoint/health/status 与 response/output/side effect。
-6. `Development Self-Test Report` 必须记录 `Report Status`、当前 task 本轮执行 `self_test_contract` 中全部 scenario 和 required gates 后的结果，并记录从本地启动或调用入口开始，到完成所有自测用例的 `Module Key Test Path`；路径必须覆盖本 task / 本模块承诺的所有可运行入口、内部关键路径、关键边界、观察点和完成证据，不能只补一句 smoke 结果，也不能复用历史 PASS、模板字段、代码阅读或无关通用 gate 作为本轮证据。
+6. `Development Self-Test Report` 必须记录 `Report Status`、当前 task 本轮执行 `self_test_contract` 中全部 scenario 和 required gates 后的结果，并记录从本地启动或调用入口开始，到完成所有自测用例的 `Module Key Test Path`；路径必须覆盖本 task / 本模块承诺的所有可运行入口、内部关键路径、关键边界、观察点和完成证据，不能只补一句 smoke 结果，也不能复用历史 PASS、模板字段、代码阅读或无关通用 gate 作为本轮证据。若 contract 包含 `module_key_test_graph` 或 `graph_required: true`，报告还必须记录实际 `Module Key Test Graph`。
 7. 测试覆盖必须列出具体测试，或明确记录覆盖缺口。
 8. 文档粒度保持在模块、子系统或核心数据流级别；不要默认按 task 建文档，也不要写成跨全项目的巨型百科。
 
@@ -65,8 +65,9 @@ implementation doc 只写长期实现事实，不写完整操作日记。对于
 - [ ] 核心数据流已说明。
 - [ ] runnable entry/exit、配置契约和 fixture/live 边界已记录，或缺失项已明确标注。
 - [ ] `Development Evidence` 已记录 `Evidence Level`、`Target Runtime Environment`、`Runnable Entry`、`Observable Exit`、`Client / Server Initialization`、`Config Contract`、`Testing Handoff Readiness`、`Known Missing Runtime Boundaries`、`Basic Self-test Evidence`，或带原因的 `Not applicable`。
-- [ ] 如果当前 task 有 `self_test_contract.status: "required"`，`Development Self-Test Report` 已记录 `Report Status`、contract source、scenario results、executed gates、Module Key Test Path、actual evidence、missing/blockers 和 Testing Handoff Readiness。
-- [ ] 如果当前 task 是 high-risk runtime/live/remote-operator 工作，implementation doc 主线只保留实现事实、`Current Operator Path` 和恢复链接，`Gate Breakdown` 已分层记录，本轮失败探索已隔离到 exploration appendix。
+- [ ] 如果当前 task 有 `self_test_contract.status: "required"`，`Development Self-Test Report` 已记录 `Report Status`、contract source、Module Application Entry、scenario results、executed gates、Module Key Test Path、Observable Exit、Current Blocker、Testing Handoff Readiness 和 Evidence Index Refs。
+- [ ] 如果 contract 包含 `module_key_test_graph` 或 `graph_required: true`，`Development Self-Test Report` 已记录实际 `Module Key Test Graph`，且图内只有路径骨架和 evidence pointer。
+- [ ] 如果当前 task 是 high-risk runtime/live/remote-operator 工作，implementation doc 主线只保留实现事实、短 `Current Operator Path`、hard constraints 和恢复链接，`Gate Breakdown` 已分层记录，本轮失败探索已隔离到 exploration appendix。
 - [ ] `business_handoff_ready` task 已记录 Testing Handoff Contract。
 - [ ] 已判断 implementation doc 的语义切片边界。
 - [ ] 方案偏移和测试覆盖已记录。